Research On 3D Indoor Scene Object Recognition Method Based On CNN

With the rapid development of smart home industry,intelligent devices have sprung up,and constantly meet the needs of people's quality of life.Multi-functional home robot has the characteristics of independence,convenience and autonomy,to a certain extent,the family plays a role in sharing housework,security protection,loved by the majority of families and respected.The identification and classification of objects has been an important research topic in the field of robot vision,especially the object recognition and classification of indoor environment is an important direction of development in recent years,the traditional identification method is very difficult to find good solutions.In this dissertation,the LCCP^[1]algorithm is used to optimize the convolution neural network technology,in the 3D indoor scene for indoor object recognition,speed up the object recognition speed,improve the accuracy of object recognition,the main work:The first processing stage,the pre segmentation method used in this dissertation is the3D information of the scene based on scene segmentation,preprocessing by obtaining the depth information of indoor scene samples in the RGB-D database,using the LCCP algorithm from the indoor scene,object of complex independent segmentation,image segmentation will be normalized to the size of the 256*256 image.Secondly,in the object recognition stage,this dissertation designed a 9 layer of the convolutional neural network to identify the indoor objects,based on classical convolutional neural network based on LeNet-5 and AlexNet structure based on the classical neural network by changing the network structure,control the convolution kernel size,select activation function and method of pool network optimization.The structure of the network consists of 4 layers:two layers,3 layers,and all the connected layers,which can effectively prevent over fitting problem by controlling the size of convolution kernel and Dropout.Finally,based on the RGB-D data set for experiment,and according to the experimental process to improve the network,through the different layers,adjusting the parameters of core network structure of different size convolution,finally get the optimal results.Compared with the traditional method of object recognition and other kinds of neural network recognition,the neural network has a higher recognition rate.